Integrating Digital Image Processing and Artificial Neural Network for Estimating Porosity from Thin Section

Porosity estimation from thin section image using digital image processing is critical for petrography study since it gives a brief description on the 2D porosity of the sample. The standard routine uses the binarization process that converts the colour (RGB) image into a binary image using pixel value treshold. The idea is that the treshold value must accomodate all the blue regions correlate to pore and turns it into white in the resulting binary image. Errors come from mis-conversion when the matrix is converted into pores and yields larger pore spaces. This error can be larger when dealing with samples having high blue variation, for example; stained carbonate or polarized microscope captured thin section. To address this issue, an integration of neural network and image processing is proposed to estimate accurately the porosity from a thin section. The neural network, based on color pattern difference between pores and matrix is assigned to segment them by addressing their original pixel value with a targeted code value, e.g.; 1 for pores and 0 for matrix. After the segmentation, pixel counting is then used to calculate the porosity. This method requires the training data which is the cropped region of pores and matrix. In the training process it uses the Lavenberg-Marquadt learning method. Synthetic samples and bonafide thin sections of sandstone and stained carbonate are used for testing. For each sample the computation time is less than 10 seconds and the error, which is caused by mis-segmentation, is less than 1%. Results from standard image processing technique are also evaluated for comparison.The algorithm can also work for the grey scale image and binary image as such it is suggested that it can work for estimating microporosity in carbonates or coal from SEM image. The proposed method brings new advance in digital petrography and can be an accurate and fast tool for porosity estimation from thin sections. Furthermore, as the neural network algorithm can also segment the multi mineral matrix present in the sample (stained carbonate thin section), the fraction of each mineral constituent can also be estimated.